Automated Modeling of Parallel Algorithms for Performance Optimization and Prediction.
Robert C. Durie, Adam W. Bojańczyk · 1997
We describe a tool for building parallel applications on distributed memory machines. The central component of this tool is the Parallel Program Factory (PPF). The input to the PPF is a dataflow graph describing the application on the level of functional blocks. Given the graph and a collection of parallel libraries, the PPF automatically selects an efficient parallel implementation of the application. A Space Time Adaptive Processing problem developed by the signal processing community is used to demonstrate the concept. 1 Introduction Processing speeds for a single processor are not sufficient to satisfy the computational demand of large signal processing algorithms. For these classes of problems, parallel processing is required to get the necessary performance. These real-time signal processing applications are usually implemented on distributed memory machines. This is because the distributed memory model exposes the cost of interprocessor communication, which is often a bo...